Researcher profile: Nataša Sladoje

Nataša is a professor in computerized image analysis. She develops algorithms which enable to efficiently and reliably, by using computers, extract and interpret information from digital images.

- We, humans, process enormous amounts of information through our visual system and we consider vision the most important of our senses. Development of methods that enable automated analysis of the huge, and rapidly growing amount of visual data which are acquired by an abundance of different sensors in a variety of application fields, is not only essential for the advancement of these fields, but is also naturally attractive to humans. We simply like to teach our computers to “see” what we can see ourselves!

Nataša started her academic career in mathematics and eventually found interest in visual data analysis.
- My background is in mathematics and my entrance point to the field of digital image analysis was via discrete geometry, which is one of the fundamental theoretical frameworks for representing and interpreting digital images. I became fascinated by the intuitiveness and interpretability of many theoretical results in discrete geometry, once they were put in the context of visual data analysis.

In this field, it is not uncommon to collaborate with researchers and professionals from many different parts of society. This is not surprising, considering how important processing of visual data has become, and how diverse this data is.
- Computerized Image Analysis is a highly interdisciplinary field. Images are signals, big data, abstract mathematical structures, and visual representations of phenomena in a variety of applications. It typically requires collaborative work of a whole bunch of experts – physicists and engineers, computer scientists, data scientists, statisticians and mathematicians, and experts with the domain knowledge – medical doctors, astronomers, archaeologists, life scientists… – to enable a computer to do what we humans can do so naturally – process and interpret visual data.

- In life sciences, powerful and complex imaging techniques can reveal a variety of properties of a specimen of interest – morphology, dynamics, function – however, most often only one such property at a time, Nataša explains.

To reach a holistic view, we have to combine multiple different techniques, including AI-based ones.
- For full understanding of the objects of interests and processes which involve them, different complementary techniques have to be combined. I am interested in the development of AI-based methods which enable reliable, interpretable and trustworthy utilization and analysis of information-rich multimodal image data, in particular in biomedicine.

- AI-based methods, and specially those based on deep learning, are nowadays dominant in the field of image analysis, due to their generally outstanding performance. They have a potential to enable integration and analysis of heterogeneous multimodal information in the most challenging scenarios, she says.
- A prerequisite for utilization of AI-supported analysis and decision systems in biomedicine and health care is interpretability of the obtained results. Medical doctors must be given ways to understand the basis for the decisions, which allows them to decide if they can trust the system. We strive for solutions based on Explainable Artificial Intelligence (XAI). Trustworthy AI-systems will open many paths towards discovery of new knowledge in science, by highlighting relevant patterns and inter-relations in the data that we were not aware of before.

Development of these methods is an active and expanding research field; Nataša's research group MIDA (Methods for Image Data Analysis) have proposed several new ways how AI can be used for improved image analysis.
- We are developing AI-based methods for multimodal image analysis, in particular in the context of cancer detection. Our methods for multimodal image alignment reach state of the art performance in several multimodal biomedical and medical application scenarios. The methods are shared publicly and used by other researchers. We are now integrating them into our developed AI-based decision-support system for oral cancer detection. We believe that by using diverse multimodal information we can increase both reliability and interpretability of the system.

The MIDA research group actively collaborates with experts within life sciences and the healthcare sector.
- Several previous and current members of MIDA, the research group which I am leading at the IT Department, work closely with collaborators from life sciences and healthcare. Together, we plan to apply our developed methods in different scenarios within biomedicine, and to evaluate their potential to advance everyday health care, in particular for improved cancer detection, but also for better understanding of the disease.

Nataša is also a group leader in the international network COMULIS.
- A large European COST network COMULIS (Correlative Multimodal Imaging in Life Sciences), where I act as a work group leader, provides a rich international interdisciplinary scientific environment which further stimulates synergies towards exciting research in life sciences and advanced computational method development.

However, that also comes with its own complexities.
- As often in highly interdisciplinary fields, a big challenge is to learn to communicate with collaborators coming from different scientific backgrounds, to find a common language across the disciplines, to understand the needs, and limitations of different methodologies involved. However, this is crucial for success, and is (once reached) also highly inspiring and rewarding.

Some final words from Nataša.
- Digital images are always only approximations of the objects and scenes from the real world. Imaging techniques are limited by the underlying physical processes, as well as optimized imaging conditions (e.g., to prevent damages that may be caused). Even though the amount of acquired data is tremendous, its informative content often appears insufficient. Scientists (and humans in general) will always wish to see further, smaller, faster events and objects – to picture the world with more details and crisper colours.

- I have worked a lot on developing methods which maximize the information extracted from available data. These include approaches to compensate for limited image resolution when performing precise measurements of objects, approaches to improve the quality of images acquired under specific constraints, and – what I am most interested in right now – approaches to meaningfully combine heterogeneous information from images acquired by different sensors, to reach new and higher levels of understanding of complex phenomena, Nataša says.

And the future?
- Considering that modern image analysis heavily relies on artificial intelligence, machine learning and data driven approaches, a very important next step is to develop methods to teach our automated systems to “reason”, and to go beyond detecting patterns (correlations) in data, towards understanding causal relations and (perhaps even) reaching truly intelligent behaviours.

- This, however, still seems to be a long-term goal, but small steps towards it are continuously being made. We are very interested in joining the effort!

Interview by Victor Kuismin, March 18, 2022